Edit Only Where Unsure: Lesion-Level Locking for Liver Tumor Mask Refinement

Automatic "refiners" are models that take a first-draft output and try to improve it. They now appear in both translation and medical image segmentation, and they share one failure mode: when the draft is already good, the refiner changes it anyway and makes it worse. In translation this is called over-correction. The field's answer has been to decide whether to edit before deciding how. A separate model estimates which parts of the draft are fine and forbids the editor from touching them. Liver tumor segmentation has the same structure. A base network outlines tumors on CT, and a refiner such as the Segment Anything Model (a general-purpose promptable segmenter) then redraws the outlines. Those redrawn outlines are known to erode edges and add blob-like artifacts. This matters because tumor outlines feed decisions. Whether a patient with hepatocellular carcinoma (HCC, the main primary liver cancer) has been "downstaged" by preoperative therapy into transplant eligibility depends on millimetre-scale diameters and lesion counts. The Milan criteria, for example, allow one tumor up to 5 cm, or up to three tumors each up to 3 cm.

We propose to carry the translation field's word-level quality gating over to 3D liver tumor segmentation. For each lesion, a reference-free quality estimate decides whether that lesion is "OK" and locked or "suspect" and open to refinement, and the refiner may only change suspect lesions. We evaluate the method as a way to prevent false stage-down calls. Our main test is a new one: we measure how often a pipeline "detects" tumor shrinkage that crosses a Milan threshold when it is shown the same scan twice under harmless perturbations. On such a pair, any change it reports is an artifact.

Core Claim

Gating refinement at the lesion level, using agreement across an ensemble of segmenters, will sharply reduce the fraction of lesions a refiner makes worse, without giving up the refiner's average gains. It will also cut the rate of spurious Milan-status changes compared with both always refining and gating whole scans. The new ingredient is neither refinement nor quality estimation, since both exist. It is feeding a spatially resolved quality estimate back into the refiner as a hard constraint, and doing so at the right granularity.

Prior Work and Where This Sits

Our method comes from translation post-editing. As translation drafts improved, the WMT post-editing findings (Findings of the WMT 2022 Shared Task on Automatic Post-Editing) reframed the problem around whether to edit at all. LangMark found that the best large language model editor won because it was conservative and left adequate translations alone. The most direct template is QE-assisted constrained decoding. It tags words OK or BAD with a quality-estimation model, forces OK words into the output through constrained beam search, and reports TER gains of 0.65 (En-De), 1.86 (En-Hi), and 1.44 (En-Mr) along with fewer deteriorations. Two of its findings shape our design. First, word-level constraints beat sub-word constraints; our analogue is lesion-level versus voxel-level locks. Second, a wrong OK tag locks an error in permanently, so we measure that failure explicitly.

Two medical papers have each taken one half of this idea. Robust, Primitive, and Unsupervised Quality Estimation for Segmentation Ensembles scores each ensemble member against a consensus map with a median-absolute-deviation threshold. Its alarm counts correlate with true Dice at r = −0.72 (glioma MRI) and r = −0.70 (lung CT). However, it only raises alarms, never touches the segmentation, and was not tested on liver. ASLseg documents over-correction by SAM on liver tumors: edge shrinkage, over-segmentation, and blobs. It adds an adaptation network trained on deliberately corrupted masks, and accepts or rejects whole pseudo-labels with a fixed Dice-agreement threshold. It reaches 74.28% Dice on LiTS with 10% labels, against 72.16% for the prior best and 76.16% fully supervised, but it works slice-wise and at the level of whole masks. Our method differs from both in the same way. We use the quality estimate spatially, lesion by lesion in 3D, as a constraint on what the refiner may change.

The clinical papers define why errors matter rather than supplying method. LiTS shows lesion segmentation remains weak: top tumor Dice is 0.674–0.739, and lesion recall is 0.554, collapsing for lesions under 10 mm. These are exactly the lesions that decide lesion counts under Milan. The systematic review of pre-transplant checkpoint-inhibitor therapy shows that downstaging-to-transplant decisions carry real cost: 19% pooled allograft rejection and 24% post-transplant recurrence. A false "downstaged" call is therefore not harmless. The finotonlimab plus bevacizumab-biosimilar trial reports a 32.8% objective response rate, meaning a sizeable minority of patients shrink on therapy and will be measured for eligibility. The hepatectomy nomogram study finds maximum diameter and non-smooth margin to be independent predictors, yet extracts them manually. Boundary fidelity is thus clinically loaded, and automating the extraction is only safe if refiners do not distort margins.

We do not claim to optimise treatment or washout timing. Nothing in the file supports modelling that, so the clinical direction enters only as an evaluation target.

Research Program

Base model and ensemble. A 3D base segmenter produces a draft tumor mask, and each connected component of that mask is treated as one lesion. An ensemble of K diverse segmenters is trained on the same labeled split. Architectures, seeds, and input phases vary; we use K = 5 by default.

Quality score. For each draft lesion we compute its agreement with the ensemble consensus: component-level Dice and 95th-percentile surface distance, plus whether the consensus contains an overlapping component at all. Following the robust quality-estimation paper, a lesion is tagged OK if its agreement exceeds median − α·MAD computed over lesions in the validation fold. Otherwise it is tagged suspect.

Refinement. The refiner is either SAM prompted from the draft or ASLseg's adaptation network. It runs only on suspect lesions, prompted by their bounding boxes. Its output is then merged under a hard constraint: voxels belonging to locked lesions, and a background margin around them, keep their draft labels. New components the refiner proposes outside any draft lesion are admitted only if the ensemble consensus also contains them. This addresses missed small lesions without letting SAM hallucinate new blobs.

Granularity variants. Mirroring the word-versus-sub-word finding, we test three lock granularities. Case-level locks accept or reject refinement for the whole scan. Lesion-level locks are the default. Voxel-level locks freeze only boundary voxels where the ensemble agrees, so the refiner may edit the rest.

Clinical readout. From every final mask we compute maximum lesion diameter and the count of lesions above a minimum size. From these we derive a Milan-style status. LiTS mixes primary and secondary liver tumors, so we treat this status as a burden-threshold proxy rather than a true HCC eligibility label.

Baselines

No refinement. The reference point is the base 3D segmenter's draft, used as is. Every other system starts from this exact draft, so any difference is attributable to refinement and gating. Beating it means higher lesion-level Dice or small-lesion recall without a higher false-change rate.

Unconstrained refinement. This is the always-edit system. The same SAM or adaptation-network refiner is applied to every lesion with identical prompts and weights. It is the direct analogue of the unconstrained post-editors that the translation work improved on. Our method must match its mean Dice within about half a point while at least halving its lesion deterioration rate. We define deterioration as the fraction of lesions whose Dice drops by more than 0.02 after refinement.

ASLseg. This is the strongest existing method on this question, because it was built specifically to contain SAM over-correction on LiTS. Its reported result is 74.28% Dice at 10% labels (ASLseg). We rerun it on the same 10% labeled split and the same test fold, with the same labeled data and GPU budget as our ensemble. Its consistency threshold gets the same number of validation tuning trials as our α. We found no reported result on Lacuna for deterioration rate or lesion-level edit behaviour on LiTS for any method, so those numbers come from our reruns. Beating ASLseg means Dice at or above its rerun value together with lower deterioration and a lower false-change rate. If we reach its Dice but not the other two, that counts as a tie.

Case-level quality gate. This baseline applies the robust quality-estimation alarm to the whole scan and refines only alarmed scans. It uses the identical ensemble, so it isolates the effect of spatial granularity.

Ablations. A random-lock ablation freezes the same fraction of lesions chosen at random, which tests whether the quality signal matters beyond the mere act of editing less. An oracle-lock ablation freezes lesions whose true Dice is already high, which bounds what better quality estimation could buy. We also run the same whole pipeline trained with 100% labels, to check that the gains are not an artifact of low-label noise.

Why not just…?

Why not just gate whole scans, as ASLseg and the robust quality-estimation paper effectively do? This is the most serious threat. If most scans are either uniformly good or uniformly bad, a case-level gate captures nearly all the benefit. Two facts argue against that. LiTS scans often contain several lesions of very different sizes, and small-lesion recall of 0.554 means good and bad lesions coexist within one scan. The translation evidence also favours finer constraints, up to the word level. Still, this is an empirical question, and the granularity comparison settles it. If case-level gating comes within noise of lesion-level gating on deterioration and false-change rate, we will report that the simpler gate suffices and drop the lesion-level claim.

Why not just train a stronger base segmenter and skip refinement? A stronger base model raises every system, including ours, and we test this with the 100%-label run. It does not remove the problem. Refiners are used precisely because base models miss small and faint lesions, and whenever a refiner is used the question of when to trust it remains. If refinement adds nothing over the base model at 100% labels, that is itself a reportable finding: gating would then only matter in the low-label regime.

Why not just use the refiner's own confidence, or test-time augmentation variance, instead of an ensemble? This is cheaper, and we include it as an alternative quality score inside the same locking scheme. The concern is that SAM tends to be confidently wrong on medical edges; the edge erosion ASLseg reports illustrates this. Independent models disagreeing is the signal the robust quality-estimation paper validated. Whichever score yields the lower lock-in error rate at matched lock fraction wins. Lock-in error rate is the fraction of locked lesions with true Dice below 0.5.

Proposed Experiments and Decisive Checks

All experiments use LiTS with 10% and 100% labeled splits and a held-out test fold. We report lesion-level and global Dice, average symmetric surface distance, lesion recall and precision overall and for lesions under 10 mm, deterioration and improvement rates, lock-in error rate, and agreement between mask-derived Milan-style status and the status derived from ground truth.

Primary check: no-change stability. Each test volume is paired with a perturbed copy of itself: resampled slice thickness, mild intensity and noise shifts, and small rigid shifts. The biology is identical, so every reported shrinkage of the largest diameter by more than 10%, and every flip in Milan-style status, is a false stage-down or false progression call. The prediction is that unconstrained refinement produces the most false flips, since each run independently re-edits edges. Lesion-level locking should produce the fewest, because agreed-upon lesions are frozen in both runs.

Falsification. If lesion-level locking does not beat both case-level gating and random locks on this false-flip rate, the core claim fails.

Secondary check: real sensitivity. We apply synthetic erosion to lesions to confirm that locking does not suppress genuine changes. For true longitudinal pre- and post-therapy validation, the file offers no public dataset. It would require an institutional cohort of patients scanned before and after downstaging therapy, which we flag as the next step rather than claim here.